PMSM transient response optimization by end-to-end optimal control

TL;DR

The paper presents an end-to-end Lipschitz-regularized ReLU RNN controller for PMSMs that directly minimizes speed transient responses via an optimal-control-inspired loss, replacing traditional PI-FOCs. By unrolling the plant with RK4 in a differentiable framework, the approach learns arbitrary current trajectories in the $i_d$–$i_q$ plane, enabling effective flux-weakening traversal and improved acceleration, especially for fast ramps, while maintaining copper losses comparable to or better than reference FW strategies. The method demonstrates robustness to parameter variations and load fluctuations in simulation, and it discusses practical considerations such as inversion modeling, iron losses, and stochastic effects as avenues for future work. Overall, the study highlights the viability of end-to-end learning for real-time PMSM control, offering a data-driven route to enhanced transient performance without hand-tuned current references.

Abstract

Speed responses of motors, especially Permanent Magnet Synchronous Motors (PMSMs), are increasing in importance for recent applications, such as electric vehicles or quadrotors. These applications require quick acceleration performance. However, commercial controllers are based mainly on Proportional-Integral (PI) controllers, which are suitable for eliminating steady-state errors but unsuitable for transient response optimization. In this paper, we replaced whole conventional controllers with an end-to-end Recurrent Neural Network (RNN) that has a regularized transition matrix. Our end-to-end controller directly minimizes the transient response time on the basis of optimal control theory. Computer-simulated results show that speed response indices improved using the RNN rather than a PI controller, while both were under comparable power losses. The current vector trajectories of the RNN showed that the RNN could automatically determine arbitrary trajectories in the flux-weakening region in accordance with an arbitrarily designed loss function. In contrast, the traditional flux-weakening methods using PI controllers have pre-determined current vector trajectories.

Figures (32)

• Figure 1: Baseline PI-FOC system. Current reference limiters and anti-windup components are added if the baseline performance, depending on the simulation condition, is not enough. Note that the proposed RNN controller replaces all of the components inside the dotted line as a black-box.
• Figure 2: Dataset definition. $N_{\rm batch}$ points in training dataset are sampled per epoch.
• Figure 3: Training loss throughout RNN training epoch
• Figure 4: Evaluation loss throughout RNN training epoch
• Figure 5: Training response indices throughout RNN training epoch